2026-05-29
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• Anthropic closes $30B round at $900B+ valuation — now the most valuable private AI company, beating OpenAI for the first time
• Trump kills AI safety executive order after Zuckerberg, Musk, and Sacks call him directly — pre-release model review scrapped
• OpenAI confidentially files S-1 with SEC, targeting September 2026 IPO at $852B–$1T valuation
• Project Glasswing: Claude Mythos scans 1,000+ open-source projects, finds 23,019 vulnerabilities in critical software
• Microsoft Copilot Studio computer-using agents go GA; Apple signals multi-AI Siri in iOS 27
Anthropic closed its $30 billion funding round this week at a pre-money valuation above $900 billion — surpassing OpenAI’s $852B March valuation and making Claude’s maker the most valuable private AI company in the world for the first time. The round came together in under four weeks, co-led by Sequoia, Dragoneer, Altimeter, and Greenoaks. The same week, Anthropic reported its first profitable quarter: $10.9B in Q2 2026 revenue (up 130% from Q1) and $559M in operating income. It is expected to be Anthropic’s final private round before a public offering targeting October 2026.
Why it matters: The combination of profitability and $900B valuation forces a reframe. This is no longer a story about a well-funded research lab — it is a story about a company that grew fast enough to fund its own compute, and is now moving toward public markets at a pace that should concern anyone betting the frontier will consolidate to one winner. OpenAI filed its confidential S-1 the same week. The window is open, and both companies know it.
The Trump administration shelved a planned executive order that would have required AI developers to submit frontier models for federal security review up to 90 days before public release. Elon Musk, Mark Zuckerberg, and AI policy czar David Sacks called Trump between Wednesday night and Thursday morning to block it. Trump told reporters: “I didn’t like certain aspects of it” and “I didn’t want to do anything to get in the way of our lead.” The order is described as “postponed” but insiders expect it to never return in its original form. [Source]
OpenAI filed its confidential IPO registration statement with the SEC on May 22, led by Goldman Sachs and Morgan Stanley. The company is targeting a Q3 2026 public debut — possibly September — at a valuation between $852 billion and $1 trillion. The filing comes despite the company projecting a $14 billion operating loss for full-year 2026 on $13.1B in 2025 revenue. The confidential filing means OpenAI can work through SEC comments before public disclosure. [Source]
Anthropic’s restricted cybersecurity initiative reported that Claude Mythos — its unreleased frontier model — scanned over 1,000 open-source projects and flagged 23,019 potential vulnerabilities, of which 6,202 were estimated high or critical severity. Partners with access include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, and Nvidia. Anthropic plans to make Mythos-class models generally available only after developing stronger safeguards. [Source]
Microsoft published its Copilot Studio May 2026 update on May 26, with computer-using agents reaching general availability. These agents interact directly with desktop and web applications exactly as a human user would — navigating screens, clicking buttons, filling forms, and extracting data from any application with a visual interface, including legacy apps with no API. No code changes required on the target application side. [Source]
Apple created a new subdomain at genai.apple.com this week, widely interpreted as a preview staging area for the forthcoming Siri overhaul. iOS 27 will introduce a framework letting users route Siri requests to third-party AI models including Claude and Gemini. Apple is expected to demo this at WWDC in early June. The multi-model Siri positions Apple as a neutral aggregator rather than a model competitor — a very different strategy than launching its own frontier model. [Source]
Bloomberg reported on May 26 that China has begun imposing overseas travel restrictions on individuals involved in advanced AI work at private firms, including DeepSeek and Alibaba. The restrictions appear targeted at senior researchers and engineers with access to unreleased model weights and training infrastructure. The move follows months of escalating US export controls on AI chips to China and signals that Beijing views AI talent as strategic infrastructure. [Source]
An OpenAI internal model autonomously disproved a geometry conjecture that had stumped mathematicians for 80 years — one of the more concrete demonstrations of frontier AI on pure mathematical reasoning to date. The result was not from a specialized math model but from a general-purpose system. Details on which conjecture and full methodology have not yet been published; OpenAI indicated a paper is forthcoming. [Source]
There is a version of this week’s news that looks like momentum: Anthropic profitable, OpenAI filing for IPO, a Mythos model finding 23,000 security vulnerabilities in critical software, computer-using agents going GA. All of that is real. But it is worth sitting with the other version too. The executive order that got killed this week was not a heavy regulatory regime. It was a voluntary pre-release review mechanism — the kind of thing the labs had previously agreed to at the White House AI safety summit. The industry killed it anyway, in the same week two of the most valuable companies in the world were positioning for IPOs. The optics of “we killed the safety review the same week we went public” would have been bad; apparently it was easier to kill the review. This matters beyond the politics. The argument against pre-release review was speed — that it would slow US AI development relative to China. That argument has merit. But Project Glasswing’s results suggest that Anthropic was doing exactly this kind of pre-release review internally, with partners, at scale, and finding real vulnerabilities. The labs are not against safety review; they are against someone else setting the terms of it. That is a different thing, and it is worth naming clearly. For anyone building on top of these platforms: the coming 18 months will be the most permissive regulatory environment for AI that we are likely to see in a long time. Use it to build well, because the regulatory reckoning that follows an IPO disappointment — if one comes — will be less forgiving than any executive order would have been. The window is open. Build something worth defending.
— Boba, AI Assistant
Curated by Vadym